BLAKE3/c/README.md

The official C implementation of BLAKE3.

Example

An example program that hashes bytes from standard input and prints the result:

#include "blake3.h"
#include <stdio.h>
#include <unistd.h>

int main() {
  // Initialize the hasher.
  blake3_hasher hasher;
  blake3_hasher_init(&hasher);

  // Read input bytes from stdin.
  unsigned char buf[65536];
  ssize_t n;
  while ((n = read(STDIN_FILENO, buf, sizeof(buf))) > 0) {
    blake3_hasher_update(&hasher, buf, n);
  }

  // Finalize the hash. BLAKE3_OUT_LEN is the default output length, 32 bytes.
  uint8_t output[BLAKE3_OUT_LEN];
  blake3_hasher_finalize(&hasher, output, BLAKE3_OUT_LEN);

  // Print the hash as hexadecimal.
  for (size_t i = 0; i < BLAKE3_OUT_LEN; i++) {
    printf("%02x", output[i]);
  }
  printf("\n");
  return 0;
}

If you save the example code above as example.c, and you're on x86_64 with a Unix-like OS, you can compile a working binary like this:

gcc -O3 -o example example.c blake3.c blake3_dispatch.c blake3_portable.c \
    blake3_sse41_x86-64_unix.S blake3_avx2_x86-64_unix.S blake3_avx512_x86-64_unix.S

API

The Struct

typedef struct {
  // private fields
} blake3_hasher;

An incremental BLAKE3 hashing state, which can accept any number of updates. This implementation doesn't allocate any heap memory, but sizeof(blake3_hasher) itself is relatively large, currently 1912 bytes on x86-64. This size can be reduced by restricting the maximum input length, as described in Section 5.4 of the BLAKE3 spec, but this implementation doesn't currently support that strategy.

Common API Functions

void blake3_hasher_init(
  blake3_hasher *self);

Initialize a blake3_hasher in the default hashing mode.

void blake3_hasher_update(
  blake3_hasher *self,
  const void *input,
  size_t input_len);

Add input to the hasher. This can be called any number of times.

void blake3_hasher_finalize(
  const blake3_hasher *self,
  uint8_t *out,
  size_t out_len);

Finalize the hasher and emit an output of any length. This doesn't modify the hasher itself, and it's possible to finalize again after adding more input. The constant BLAKE3_OUT_LEN provides the default output length, 32 bytes.

Less Common API Functions

void blake3_hasher_init_keyed(
  blake3_hasher *self,
  const uint8_t key[BLAKE3_KEY_LEN]);

Initialize a blake3_hasher in the keyed hashing mode. The key must be exactly 32 bytes.

void blake3_hasher_init_derive_key_raw(
  blake3_hasher *self,
  const void *context,
  size_t context_len);

Initialize a blake3_hasher in the key derivation mode. Key material is to be given as input after initialization, using blake3_hasher_update. The key derivation context should follow the Key Derivation Context Guidelines described below. context_len indicates the size of context in bytes.

void blake3_hasher_init_derive_key(
  blake3_hasher *self,
  const char *context);

Similar to blake3_hasher_init_derive_key_raw, except it takes the key derivation context as a null-terminated C string.

This function is offered as a convenience. It is recommended to use this function when giving a literal, hardcoded C string as parameter.

Notice that contrary to blake3_hasher_init_derive_key_raw, this function cannot accept contexts containing the byte 0x00 except as a the terminating byte. For this reason, blake3_hasher_init_derive_key_raw is preferred in more general contexts, such as when implementing bindings to this C library.

void blake3_hasher_finalize_seek(
  const blake3_hasher *self,
  uint64_t seek,
  uint8_t *out,
  size_t out_len);

The same as blake3_hasher_finalize, but with an additional seek parameter for the starting byte position in the output stream. To efficiently stream a large output without allocating memory, call this function in a loop, incrementing seek by the output length each time.

Key Derivation Context Guidelines

The key derivation context should uniquely describe the application, place and purpose of the derivation.

The context should be statically known, hardcoded, globally unique, and application-specific.

The context should not depend on any dynamic input such as salts, nonces, or identifiers read from a database at runtime.

A good format for the context string is:

[application] [commit timestamp] [purpose]

For example:

example.com 2019-12-25 16:18:03 session tokens v1

It's recommended that the context string consists of ASCII bytes containing only alphanumeric characters, whitespace and punctuation. However, any bytes are acceptable as long as they satisfy the static constraints described above.

Building

This implementation is just C and assembly files. It doesn't include a public-facing build system. (The Makefile in this directory is only for testing.) Instead, the intention is that you can include these files in whatever build system you're already using. This section describes the commands your build system should execute, or which you can execute by hand. Note that these steps may change in future versions.

x86

Dynamic dispatch is enabled by default on x86. The implementation will query the CPU at runtime to detect SIMD support, and it will use the widest instruction set available. By default, blake3_dispatch.c expects to be linked with code for four different instruction sets: portable C, SSE4.1, AVX2, and AVX-512.

For each of the x86 SIMD instruction sets, two versions are available, one in assembly (with three flavors: Unix, Windows MSVC, and Windows GNU) and one using C intrinsics. The assembly versions are generally preferred: they perform better, they perform more consistently across different compilers, and they build more quickly. On the other hand, the assembly versions are x86_64-only, and you need to select the right flavor for your target platform.

Here's an example of building a shared library on x86_64 Linux using the assembly implementations:

gcc -shared -O3 -o libblake3.so blake3.c blake3_dispatch.c blake3_portable.c \
    blake3_sse41_x86-64_unix.S blake3_avx2_x86-64_unix.S blake3_avx512_x86-64_unix.S

When building the intrinsics-based implementations, you need to build each implementation separately, with the corresponding instruction set explicitly enabled in the compiler. Here's the same shared library using the intrinsics-based implementations:

gcc -c -fPIC -O3 -msse4.1 blake3_sse41.c -o blake3_sse41.o
gcc -c -fPIC -O3 -mavx2 blake3_avx2.c -o blake3_avx2.o
gcc -c -fPIC -O3 -mavx512f -mavx512vl blake3_avx512.c -o blake3_avx512.o
gcc -shared -O3 -o libblake3.so blake3.c blake3_dispatch.c blake3_portable.c \
    blake3_avx2.o blake3_avx512.o blake3_sse41.o

Note above that building blake3_avx512.c requires both -mavx512f and -mavx512vl under GCC and Clang, as shown above. Under MSVC, the single /arch:AVX512 flag is sufficient. The MSVC equivalent of -mavx2 is /arch:AVX2. MSVC enables SSE4.1 by defaut, and it doesn't have a corresponding flag.

If you want to omit SIMD code on x86, you need to explicitly disable each instruction set. Here's an example of building a shared library on x86 with only portable code:

gcc -shared -O3 -o libblake3.so -DBLAKE3_NO_SSE41 -DBLAKE3_NO_AVX2 -DBLAKE3_NO_AVX512 \
    blake3.c blake3_dispatch.c blake3_portable.c

ARM NEON

The NEON implementation is not enabled by default on ARM, since not all ARM targets support it. To enable it, set BLAKE3_USE_NEON=1. Here's an example of building a shared library on ARM Linux with NEON support:

gcc -shared -O3 -o libblake3.so -DBLAKE3_USE_NEON blake3.c blake3_dispatch.c \
    blake3_portable.c blake3_neon.c

Note that on some targets (ARMv7 in particular), extra flags may be required to activate NEON support in the compiler. If you see an error like...

/usr/lib/gcc/armv7l-unknown-linux-gnueabihf/9.2.0/include/arm_neon.h:635:1: error: inlining failed
in call to always_inline ‘vaddq_u32’: target specific option mismatch

...then you may need to add something like -mfpu=neon-vfpv4 -mfloat-abi=hard.

Other Platforms

The portable implementation should work on most other architectures. For example:

gcc -shared -O3 -o libblake3.so blake3.c blake3_dispatch.c blake3_portable.c

Differences from the Rust Implementation

The single-threaded Rust and C implementations use the same algorithms, and their performance is the same if you use the assembly implementations or if you compile the intrinsics-based implementations with Clang. (Both Clang and rustc are LLVM-based.)

The C implementation doesn't currently include any multithreading optimizations. OpenMP support or similar might be added in the future.